Pansharpening via Deep Guided Filtering Networkijipvc.org/article/IJIPVCV5I101.pdf · 2018. 5....

International Journal of Image Processing and Visual Communication

ISSN (Online) 2319-1724 : Volume 5 , Issue 1 , April 2018

http://www.ijipvc.org/ijipvcv5i101.html 1

Pansharpening via Deep Guided Filtering Network Ahmad Alsmadi#, Shuyuan Yang*, Kai Zhang#

Dep. of Electrical Engineering, Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education of China

Xi’an, China

[email protected], [email protected], [email protected]

Abstract— In order to enhance spatial details and reduce spectral

distortions in the fused image, in this paper, we introduce a simple

pan-sharpening way based on deep guided filtering network, our

proposed method was designed by combining the deep guided

filtering network with intensity hue saturation(IHS) transform.

First, by using the intensity component(I) which is calculated from

the up-sampled MS image via IHS transform; second, a deep

guided filtering network is constructed to filter the

panchromatic(PAN) image to extract a semantic detail map from

the panchromatic image; third, the details are injected into every

band of the up-sampled multi-spectral image to acquire the fused

image by an adaptive injection, which will aid in controlling the

quantity of the injected details. We empirically evaluated the

proposed method under a several data-sets from the GEO-2,

IKONOS- 2, and GF-2 satellites, and the results are compared

with those of its counterparts. The results demonstrate that the

proposed method can well-preserve spatial and spectral

information, in both subjective and objective aspects. The results

show that the proposed approach can acquire fusion results and

outperform the other methods in-indices of SAM, RMSE,

ERGAS, SAM, CC, UIQI and Q4.

Keywords— Pansharpening, intensity-hue-saturation (IHS)

transform, guided filtering network.

I. INTRODUCTION

The target of image fusion is to generate a single image by

combining relevant information from two or more images,

therefore, satellites-remote-sensing image fusion has been a hot

research topic of image fusion [1]. There are many commercial

satellites such as GEO-2, IKONOS-2, and GF-2, that provides

a high-spatial/low-spectral resolution panchromatic (PAN)

image and a high-spectral/low-spatial resolution multi-spectral

(MS) image concurrently. To recover this problem, there is an

effective way to combine a multi-spectral and a panchromatic

images, which is called remote-sensing image fusion, intended

for a result in a single image containing both a higher

spatial/spectral resolutions [2]. Today the fused remote sensing

image is a great significance technique and is used in providing

the imagery seen in the common Google-Maps/Earth and

Microsoft-Bing-Maps products. There are also some

applications within the field of remote-sensing that benefit from

pan-sharpened imagery, such as change-detection,

classification [3], vegetation identification, and lithology-

analysis [4].

In the last two decades, many fusion approaches have been

published to improve the spatial-resolution of the LR

multispectral image. The main goal of these approaches is to

inject the spatial-detail-information, which is extracted from

the panchromatic image into the multi-spectral image within a

predefined process, which is the injection gains. However, to

improve the spatial resolution of LRM band, there are two

common methods, which are the component-substitution (CS)

based method and the multi-resolution-analysis (MRA) based

method [5]. The component-substitution-based method is most

commonly used in remote-sensing image fusion, such as (IHS)

transform [6]-[7], Gram-Schmidt (GS) [8] and principal

component-analysis (PCA) [9]-[10]. The fusion images during

this method suffer from spectral distortions.

In the multi-resolution-analysis (MRA) based method, such as

the Laplacian-pyramid (LP) [11] and the “`a trous” wavelet

transform (ATWT) [12]. In order, the MRA based method can

provide more accurate spectral information but suffer from

spatial distortions. In the predefined process, the quantity of

panchromatic image spatial-detail-information that is injected

into the multi-spectral image is determined by the injection

gains, which affects the fusion result immediately [13].

Recently, Remote-Sensing-Image-Fusion Based on Adaptive

IHS and Multi-scale Guided Filter (AIHS) has been proposed

in [14]. The method in [14] takes advantage of the multi-scale

form of filtering, in order to extract sufficient details from the

panchromatic image. Hence, the obtained fused results through

this method are impressive. They used the panchromatic image

as the input image of a guided filter and the (I) component of

the multi-spectral image was used as the guided image at every

stage. Li et al. [15] proposed a multi-stage guided filter based

pan-sharpening method (MGF). They selected high-spatial

resolution panchromatic image as the guidance in transferring

the structures into the up-scaled MS images and smoothing the

PAN image itself. Qi et al. [16] also proposed multi-stage

guided filtering based on hyperspherical color transform

(HCS), they used the (I) component of the multi-spectral image

as the input of the guided filter, which was extracted from the

hyperspherical color transformation (HCT), and the PAN

image was used as the guidance-image at the first stage, but in

the second stage, at the same time, they made the PAN image

as the guidance and the filtering input. This method obtains the

fused image with a high-spatial-resolution, but the obvious

color distortions often occur in the sharpening process. Note

that, this method is more effective for WorldView-2 satellite

imagery, which utilizes more bands.

In this paper, we propose a novel pan-sharpening approach. By

using the guided filter, the approximation version of the




panchromatic image is obtained, in which the input image is the

panchromatic image in the first stage of guided filter, and the

(I) component of the multi-spectral image is used as the

guidance image, because the (I) component change has a bit-

effect on the spectral-information and due to its ease of usage.

In the second stage, the opposite of the aforementioned steps

will be executed. Moreover, to control the detail information

injected into the multi-spectral image, an adaptive improved

injection gains procedure is designed. We employ a guided

filter to filter the panchromatic image and the (I) component of

the multispectral image to obtain the approximation low-

resolution image in the first stage and the second stage,

respectively. The guided filter calculates the filtering output by

taking into consideration the content of a guidance-image, as it

can transfer the structures of the guidance-image to the filtering

output [17]-[18]. We use a multi-stage strategy, in order to be

able to extract more detail information. Hence, the intensity

component is obtained by using linear weighting. Enhancing

the-efficiency of our method can be achieved by the improving

the injection gains that are being obtained from both of the

panchromatic and each band of the multi-spectral images.

Super resolution(SR) reconstruction was accomplished by

employing an iterative-back-projection (IBP), which is become

a known and computationally efficient method for improving

the spatial-resolution of the image [19]. However, IBP

algorithm has some presenting limitations, like ringing artifacts

in the strong edge-area of the image [20]. In this method, the

IBP is being used for each band in the MS image.

Our fusion method produces a higher spatial and spectral

resolution with minimum-spectral-distortion for the fused

image among of existing methods. The contributions of our

method are synthesis, in which the (I) component is obtained

from the upsampled MS image, a multi-stage guided filter

procedure to filter the panchromatic image, which extracts

more detail Information, as a semantic detail map is injected

into each band to obtain the fused image by a model-based

procedure, then, it makes an iterative back projection (IBP) for

each band. For both of the subjective and the objective

evaluations, the experimental results indicate that our method

has best results among of other existing methods.

The rest of this paper is structured into five sections. In Section

II, briefly reviews the guided filter. In Section III our proposed

fusion method is described. We apply our proposed method on

different data-sets and compare our fusion results with some

other existing method in Section IV. In Section V, a conclusion

is presented.

II. GUIDED FILTER

As we know, the edge preserving filters have been recently a

hot topic in image processing, which has received a great

attention from the research community. In order to avoid

ringing artifacts, there are some of edge preserving smoothing

filters such as guided filter [21], weighted-least-squares [22],

and bilateral-filter [23]. Currently, the fastest edge preserving

filter is the guided filter. It is useful for detail enhancement,

image matting/feathering, and dehazing. The image guided

filter function implements edge preserving smoothing on an

image, by using the content of a second image, which is being

used as a guided image, in order to effect the filtering.

Furthermore, the guided image could be the same image. Like

other filtering processes, the guided filter is a neighborhood

process, but takes consideration while computing the value of

the output pixel; the statistics of a region in the corresponding

spatial neighborhood in the guidance-image [24]. The guided

filter can avoid gradient reversal artifacts. In this paper, the

guided filter is employed for pan-sharpening. It is worth to

mention; the guided filter is a local linear model between

filtering output image 𝑂 and the guidance image 𝑌 .

Theoretically, by using a guided image 𝑌 in filtering the input

image I, in order to get the output image 𝑂 , moreover, the

output image 𝑂 has the structures of the guidance-image, and it

can maintain the main information of the input image.

Assuming that 𝑂 is a linear transform of 𝑌 in a window 𝑤𝑘

centered at a pixel 𝑘 , it can be calculated by the following

equation:

𝑂𝑖 = 𝑎𝑘𝑌𝑖 + 𝑏𝑘∀𝑖 ∈ 𝑤𝑘 (1)

where 𝑖 is the pixel index, and 𝑤𝑘 is a squared window of size

(2r +1) × (2r +1). The following equation will be used to

represent the guided filtering operation for this paper:

𝑂 = 𝐺𝑟,𝜁(𝐼, 𝑌) (2)

where 𝐺 denotes the guided filter function, the parameters 𝑟

and 𝜁 are the inputs of guided filter, which are determined the

size and the blur degree of the guided filter, respectively. Note

that, 𝐼 and 𝑌 are the input and the guidance images,

respectively.

III. PROPOSED METHOD

The proposed pan-sharpening method is shown in Fig. 1. It is

based on the component-substitution (CS) method and the

multiresolution-analysis (MRA) method, which use the popular

edge-preserving guided filter and iterative back projection

(IBP) in injecting the extracted spatial detail information from

the PAN image into the up-sampled MS image. Our method

includes two main steps: Step1) using the PAN image to extract

the semantic detail map; Step2) computing the injection gain

for each band.

A. Extracting a semantic detail map from the panchromatic

Image

In order to acquire the spatial-information, the conventional

methods estimate the variation between the panchromatic

image and the intensity component of multi-spectral image or

any type of a low-pass filtered panchromatic image. However,

these methods suffer from the spectral-distortion because the

spatial information will blend with low-frequency components

[22]. We proposed a simple way to extract the semantic detailed




map by estimating the variation between the panchromatic

image and the multi-stage guided filter, which aids in

overcoming the aforementioned problem. The approximation

image of the input image is the result of the guided filter.

Hence, the semantic detailed map of the input image is the

variation between the input image and the approximation

image. In this paper, we present a multi-stage guided filter, in

which the I component is used as the guidance-image and the

panchromatic image as the input image in the first stage (s=1)

of guided filter. In the other stages of our proposed method, the

panchromatic image is used as the guidance-image and the

result of previous stage is used as the input image. The

procedure of the first stage can be formulated as follows:

𝐺𝐹(𝑃𝐴𝑁) = 𝐺𝑟, 𝜁(𝑃𝐴𝑁, 𝐼) (3)

Fig. 1 Flowchart of our proposed method.

where 𝐺𝐹 denotes the guided filter output. However, for 𝑠𝑡ℎ

stage (𝑠 > 1) , the procedure of the guided filter can be

formulated as

𝐺𝐹𝑠(𝑃𝐴𝑁) = 𝐺𝑟, 𝜁(𝑃𝐴𝑁(𝑠−1), 𝑃𝐴𝑁) (4)

where 𝑃𝐴𝑁(𝑠−1) is the approximation layer and (𝑠 − 1)𝑡ℎ

represents the index level of guided filtered output. The spatial

detail 𝑃𝐴𝑁𝐷𝑠 for the 𝑠𝑡ℎ stage can be shown as

𝑃𝐴𝑁𝐷𝑠 = 𝐺𝐹(𝑠−1)(𝑃𝐴𝑁) − 𝐺𝐹𝑠(𝑃𝐴𝑁) (5)

Note that, when s=1 the original PAN image is 𝐺𝐹(𝑠−1)(𝑃𝐴𝑁).

B. Computing the injection gain for each band The injection gains are defined by the quantity of the

panchromatic image spatial detail information, which is being

injected into the multi-spectral image. Consequently, the

injection gain has been shown to be critical in the spatial detail

injection model. Therefore, if the quantity of the spatial detail

is insufficient, the fused image will not reach its desired limits

and if the quantity of the spatial detail was very high, the fused

result will suffer from spectral distortion, which will cause

redundant information. However, each band in MS image has

different characteristics, cause the variety of the spectral

reflectance, therefore, each band has different trade-off

parameter β.

Accordingly, in [14], the trade-off parameters can be generated

adaptively, by resolving the next optimization problem:

min𝛽1,...,𝛽𝑛

∥ 𝑤𝑝 − ∑𝑛𝑖=1 𝛽𝑊𝑀𝑖 ∥2 𝑠. 𝑡. 𝛽1 ≥ 0, . . . , 𝛽𝑛 ≥ 0 (6)

where 𝑛 represents the number of the multi-spectral band and

𝛽𝑖 represents the 𝑖𝑡ℎ band trade-off parameter of the MS

image’s. Thenceforward, we can acquire the improved edge-

detecting-weighting-matrix as

𝑔𝑖 =𝑀𝑖

1/𝑛 ∑𝑛𝑖=1 𝑀𝑖

((1 − 𝛽𝑖)𝑤𝑝 + 𝛽𝑖𝑤𝑀𝑖) (7)

where 𝑤𝐴 represents the edge-detecting-weighting-matrix of

image A

𝑤𝐴 = 𝑒𝑥𝑝(−𝜆

|∇𝐴|4+𝜀) (8)

where ∇𝐴 denotes a gradient of the image A and 𝜆 and 𝜀 are the

tuning parameters. Fig. 2 illustrates the following procedure

obtaining the injection gain.




Fig. 2 The whole procedure of obtaining the injection gain.

Now to obtain the 𝐹 according to Fig. 1, the total semantic

details of the panchromatic image will inject into the up-

sampled multi-spectral image as formulated in Equation (9):

𝐹𝑖 = 𝑀𝑖 + 𝑔𝑖 ∑𝑘𝑠=1 𝑃𝐴𝑁𝐷

𝑠 (9)

where 𝐾 is the number of stages for the guided filter.

Then using IBP for each MS band by considering 𝐹𝑖 as LR

image and the up-sampled MS image as HR image to improve

the spatial-resolution of the fusion result. To improve more

spatial detail, we present multiple filter stages, but using an

excessive filter stage will suffer from some problems in the

fused image such as redundant information and produce

artifacts while increasing the computational cost. Fig. 3, shows

the fusion results of GEO-2 images with different guided filter

stages. The normalized results were obtained with respect to the

different filter stage as shown in Fig. 5. According to Fig. 5, the

best-fused results can be seen in stages 1 and 2.

Fig. 3 Illustration fusion results of GEO-2 images with different guided filter

stages. Figures (a) - (d) denote the filter stages 1 to 4, respectively.

IV. EXPERIMENTAL STUDY AND ANALYSIS

To evaluate the performance of our method, we consider

Wald’s protocol, which states that a fused image should be have

a close similarity to the image that the corresponding sensor

would observe at the highest spatial resolution [25]. We are

selected the original multi-spectral images as the true images to

be compared with the fusion results, therefore, the experiments

are accomplished on the degraded images. The parameter

settings, which is used in the proposed method are discussed.

To show the behavior of the proposed fusion method three data

sets i.e., IKONOS-2, GEO-2, and GF-2 are used

experimentally. After the subjective evaluation should be

appraised the performance of each fusion method

quantitatively. Hence, six typical evaluation metrics, namely,

the correlation coefficient (CC), the root mean square-error

(RMSE), universal-image-quality indexes (UIQI), the erreur-

relative-global-adimensionnelle de synth`ese (ERGAS), a

quaternion-based coefficient (Q4) index, and the spectral angle-

mapper (SAM), are used.

A. Parameters settings

The inputs of the guided filter are critical factors that must be

obtained. Moreover, an excessive number of iterations of IBP

may lose some part from the fused image. Note that, in our

proposed method, the Num-of-iterations is fixed to 5. The

experiments in this section are performed on the GEO-2 images.

Fig. 4. shows the normalized results with respect to different

input parameters of the guided filter, which are the square

window of radius r and the regularization parameter 𝜁. There

are three cases of interest: a) r=2 and 𝜁 = 0. 12; b) r=4 and 𝜁 =0. 22; c) r=8 and 𝜁 = 0. 42. Note that, the quality indicators are

calculated and some indexes are normalized to [0 1]. The better-

fused results have larger CC, UIQI, and 𝑄4, and smaller RMSE,

SAM, and ERGAS. Case (a) is shown in Fig. 4, where the fused

result achieves the optimal effect. Note that, the authors in [15]

fixed the parameters r=2 and 𝜁 = 0.001,0.8 in the first and the

second stage, respectively. Moreover, the authors in [16] fixed

the parameters r=4 and 𝜁 = 0.1 × 𝑛𝑏𝑖𝑡𝑠 , where nbits is the

radiometric resolution of each image.

Fig. 4 Performance of our fusion method with respect to different input

guided filter parameters.

In this paper, the rest of parameters are set as follows, the multi-

spectral image is matched to the same size of the panchromatic




image, the parameters 𝜆 = 10−9 and 𝜀 = 10−9 are set as in

[14].

We compared our proposed fusion method with several typical

fusion methods: the AIHS method [14]; the (MGF) method [15];

the (HCS) method [16].

B. Experiments with IKONOS-2 Data

The IKONOS-2 system simultaneously offers a four-band MS

image with 4m resolution and a single-band PAN image with

1m resolution. The results for IKONOS-2 are shown in Fig. 6.

The size of the PAN image and the MS image in this

experiment are 2048 × 2048 and 512 × 512, respectively. In

this study, by using the protocol in [25], we degrade the

original-MS and PAN images by a factor of 4, then we utilize a

simulated IKONOS-2 LR-MS image with the size of 128 × 128

and a corresponding panchromatic image with the size of 512

× 512. The original multi-spectral image with 4m resolution is

used as the true-image shown in Fig. 6 (a), in order to perform

comparative studies between the true-image and the fusion

result.

Fig. 5 Performance of our fusion method with respect to different guided filter stages.

The results of the different fusion methods are shown in Fig.

6(b)-(f). We cut a sub-image, which is located at the right-

bottom corner of the fused image, in order to demonstrate the

fusion results more clearly. Visually, spectral distortions from

MGF method (i.e. shown in Fig. 6(b)) are obvious as color

changes severely in the red corner. The fused result obtained by

the HCS method suffers from serious spectral distortion when

compared with all the other methods. The AIHS method has

improved, but the effect is not significant. To sum up, our

method outperforms the spectral and the spatial resolution. The

quantitative assessment indices results are listed in Table I,

where the best results of every quality indices are clearly

noticeable by font bold. Here, it can be clearly observed that

our method performs the superior results in the six evaluation

indexes when compared to the other methods.

C. Experiments with GEO-2 Data

In this section, we analyze another group of experimental

results of image fusion with GEO-2 data to further clarify the

performance of our fusion approach. The GEO-2 data set

provides a four band 2.8m resolution multi-spectral image and

a 0.7m resolution PAN image. The results for GEO-2 are

presented in Fig. 7. The size of the panchromatic and the multi-

spectral images in this experiment are 512 × 512 and 512 ×

512, respectively. In this study, by using the protocol in [25],

we degrade the original multi-spectral image by a factor of 4,

then we utilize a simulated GEO-2 LR-MS image with the size

of 64 × 64 and a corresponding PAN image with the size of

512 × 512. Fig. 7 (a) the original multi-spectral image with a

2.8m resolution which is used as the true-image to compare

between the fusion result.

The results of the different fusion methods are shown in Fig. 7

(b)-(f). Visually, the MGF method obtains MS images with

high-spatial-resolution, but the obvious color distortions occur

in the playground. Spectral distortions from HCS method (i.e.

shown in Fig. 7(c)) are clearly shown as color changes in the

high reflection areas, such as, the green areas.

Fig. 6 (a) Original Multi-spectral image. (b) MGF approach. (c) HCS

approach. (d) AIHS approach. (e) First stage proposed method. (f) Second stage proposed method.




TABLE I

RESULTS COMPARISON OF OUR PROPOSED METHOD WITH

SOME EXISTING APPROACHES ON IKONOS-2 DATA-SET

The AIHS method has improved, but, it changes the color of

some parts of the fused image when compared with the

reference-image such as the playground. Fig. 7(e)-(f) verify that

the proposed method results are more similar to the reference-

image than the results of the other methods. The numerical

experiments of the fused methods in Fig. 7 are listed in Table

II. From Table II, it can be clearly observed that our fusion

results have the best values for most of the quality indices.

D. Experiments with GF-2 Data

In this part, we study and analyze the experimental-results on

GF-2 data set. The GF-2 satellite captures 4m multi-spectral

images, i.e., red, green, blue, NIR images, and 1m

panchromatic image. In this experiment, the size of the

panchromatic and the multi-spectral images are 1024 × 1024

and 256 × 256, respectively. As in the IKONOS-2 experiment,

we spatially degrade the GF-2 data, then we utilize a simulated

GF-2 LR multi-spectral image with the size of 64 × 64 and a

corresponding panchromatic image with the size of 256 × 256.

The results for GF-2 are presented in Fig. 8. The original-MS

image with a 4m resolution is used as the true-image, in order

to compare it with the fusion result is shown in Fig. 8(a).

The fusion results are compared with the original multi-spectral

image, in order to obtain the desired objective measures, as

shown in Fig. 8(b)-(f). By comparing the fusion results with the

reference multi-spectral image visually, it can be observed that

the fusion result of the HCS method suffers from a high-

spectral-distortion. However, the spatial details in some areas


approach. (d) AIHS approach. (e) First stage proposed method. (f) Second

stage proposed method.

are lost. The result obtained from the MGF method suffers from

spectral distortion while losing some spatial details. As shown

in Fig. 8(d), AIHS performs the fused result with well preserved

spectral and spatial details. To sum up, our proposed method

results is a much better result and a more similar image to the

reference-image compared with the other methods.

Method CC RMSE ERGAS SAM UIQI Q4

MGF 0.9396 0.9168 0.9168 0.9422 0.9366 0.9429

HCS 21.9473 27.8220 25.6520 21.3698 22.7985 21.4905

AIHS 0.7502 0.7223 0.7368 0.7570 0.7558 0.7661

Proposed

𝟏𝒔𝒕 stage

0.9495 19.9996 2.9292 5.7394 0.9465 0.7728

Proposed

𝟐𝒏𝒅 stage

0.9476 20.4610 2.9958 5.7643 0.9449 0.7731





approach. (d) AIHS approach. (e) First stage proposed method. (f) Second

stage proposed method.

Table III reports the quantitative assessments indexes

calculated for GF-2 data set. From Table III, it can be clearly

observed that our method outperforms all the other methods.

V. CONCLUSION

In this paper, a novel image fusion method based on an IHS

injection model and a multi-stage guided filter is presented. In

the first stage of the guided filter we used the 𝐼 component,

which is acquired from the up-sampled multi-spectral image as

the input image and the panchromatic image as the guidance-

image, but in the second stage we used the panchromatic image

as the guidance-image and the result of the first stage as the

guidance-image, then the total semantic details map, which are

extracted by the multi-stage guided filter and were injected into

the up-sampled MS-image. Our contribution here is threefold.

Firstly, the 𝐼 component is acquired from the up-sampled MS-

image. Secondly, a multi-stage guided filter strategy was used

in filtering the panchromatic image, in order to acquire more

detail information. Third, the total semantic details map is

injected into every band of the up-sampled multi-spectral image

to acquire the fusion result, then, it makes an iterative back

projection (IBP) for each band. Our method is tested on three

degraded data-sets i.e., IKONOS-2, GEO-2, and GF-2, and

compared with three existing methods. The empirical-results

show that our fusion method is more robust in an image-

registration among of the other methods. Furthermore, the

proposed method is highly-effective among of the other

methods.

TABLE II


SOME EXISTING APPROACHES ON GEO-2 DATA-SET

TABLE III


SOME EXISTING APPROACHES ON GF-2 DATA-SET

REFERENCES [1] M. B. Haghighat, A. Aghagolzadeh, and H. Seyedarabi, “A non-

reference image fusion metric based on mutual information of image features,”Computers & Electrical Engineering, vol.37, no. 5, pp. 744-

756, Sep. 2011.

[2] M. Choi, R. Y. Kim, M. R. Nam, H. O. Kim, “Fusion of multispectral and panchromatic satellite images using the curvelet transform,” IEEE

Geoscience and Remote Sensing Letters, vol. 2, n. 2, pp. 136-140, Apr.

2005. [3] M. Vakalopoulou, K. Karantzalos, N. Komodakis, N. Paragios, “Graph-

Based Registration, Change Detection, and Classification in Very High Resolution Multitemporal Remote Sensing Data,” IEEE Journal of


MGF 0.8165 41.0929 10.4306 7.9985 0.8163 0.4623

HCS 0.4262 66.8888 17.4652 26.9324 0.4224 0.3369

AIHS 0.8061 43.5032 10.8956 5.8268 0.8084 0.4616

Proposed

𝟏𝒔𝒕 stage

0.8439 38.0223 9.5324 5.2947 0.8437 0.5091

Proposed

𝟐𝒏𝒅 stage

0.8461 37.97 9.5127 5.3391 0.8458 0.5190


MGF 0.9480 17.4182 4.1219 6.2736 0.9476 0.7619

HCS 0.7755 34.3642 10.2706 15.1813 0.7587 0.6646

AIHS 0.9575 16.2179 3.7496 4.6017 0.9571 0.7609

Proposed

𝟏𝒔𝒕 stage

0.9656 14.1916 3.3016 4.4401 0.9650 0.7642

Proposed

𝟐𝒏𝒅 stage

0.9597 15.4653 3.5877 4.5087 0.9596 0.7616




Selected Topics in Applied Earth Observations and Remote Sensing, vol.

9, no. 7, pp. 2940-2951, Jul. 2016. [4] W. Dong, X. Lin, X. Li, Z. Li, “A bidimensional empirical mode

decomposition method for fusion of multispectral and panchromatic

remote sensing images,” Remote Sensing vol. 6, no. 9, pp.8446-8467, Sep. 2014.

[5] M. Ghahremani, H. Ghassemian,“A compressed-sensing-based

pansharpening method for spectral distortion reduction,” IEEE Transactions on Geoscience and Remote Sensing, vol. 54, no. 4, pp.

2194-2206, Apr. 2016.

[6] T. M. Tu, S. C. Su, H. C. Shyu, and P. S. Huang, “A new look at IHS-like image fusion methods,” Information Fusion, vol. 2, no. 3, pp. 117–

186, Sep. 2001.

[7] T. M. Tu, P. S. Huang, C. L. Hung, C. P. Chang, “A fast intensity-huesaturation fusion technique with spectral adjustment for IKONOS

imagery,” IEEE Geoscience and Remote sensing letters, vol. 1, no. 4, pp.

309-312, Oct. 2004. [8] C. Laben and B. Brower, A. Craig, and V. Bernard, “Process for

enhacing the spatial resolution of multispectral imagery using pan-

sharpening,” United States patent US, 6 011 875, Jan. 2000. [9] [9] P. S. Chavez and A. Y. Kwarteng, “Extracting spectral contrast in

landsat thematic mapper image data using selective principal component

analysis,” Photogramm. Eng. Remote Sens., vol. 55, no. 3, pp. 339–348, 1989.

[10] V. P. Shah, N. Younan, and R. L. King, “An efficient pan-sharpening

method via a combined adaptative-PCA approach and contourlets,” IEEE Trans. Geosci. and Remote Sens., vol. 56, no. 5, pp. 1323–1335,

May. 2008.

[11] P. Burt and E. Adelson, “The Laplacian pyramid as a compact image code,” IEEE Transactions on communications, vol. COM-31, no. 4, pp.

532–540, Apr. 1983.

[12] J. Nunez, X. Otazu, O. Fors, A. Prades, V. Pala, R. Arbiol, “Multiresolution-based image fusion with additive wavelet

decomposition,” IEEE Transactions on Geoscience and Remote sensing, vol. 37, no.3, pp. 1204-1211, May. 1999.

[13] G. Vivone, L. Alparone, J. Chanussot, M. Dalla, A. Garzelli, G.

Licciardi, R. Restaino, L. Wald, “A critical comparison among pansharpening algorithms,” IEEE Transactions on Geoscience and

Remote Sensing, vol. 53, no. 5, pp. 2565-2586, May. 2015.

[14] Y. Yang, W. Wan, S. Huang, F. Yuan, S. Yang, Y. Que, “Remote Sensing Image Fusion Based on Adaptive IHS and Multiscale Guided

Filter,” IEEE Access, vol. 4, pp. 4573-4582, Sep. 2016.

[15] X. Li, W. Qi, S. Yue, “An Effective Pansharpening Method Based on Guided Filtering,” Industrial Electronics and Applications (ICIEA),

IEEE, pp. 534-538, Jun. 2016.

[16] W. Qi, X. Li, S. Yue, “A guided Filtering and HCT Integrated Pansharpening Method For WORLDVIEW-2 Satellite Images,”

Geoscience and Remote Sensing Symposium (IGARSS), IEEE

International, pp. 7272-7275, Jul. 2016. [17] W. Zhao, Q. Dai, Y. Zheng, L. Wang, “A new pansharpen method based

on guided image filtering: A case study over Gaofen-2 imagery,”

Geoscience and Remote Sensing Symposium (IGARSS), IEEE International, pp.3766-3769, Jul. 2016.

[18] X. Yin, J. Hu, J. Xu. “Contactless fingerprint enhancement via intrinsic

image decomposition and guided image filtering,” Industrial Electronics and Applications (ICIEA), 2016 IEEE 11th Conference on. IEEE, pp.

144-149, Jun. 2016.

[19] H. Li, K. M. Lam, “Guided iterative back-projection scheme for single image super-resolution,” Global High Tech Congress on Electronics

(GHTCE), 2013 IEEE. IEEE, pp. 175-180, Nov. 2013.

[20] R. Nayak, S. Harshavardhan, D. Patra, “Morphology based iterative back projection for super-resolution reconstruction of image,” Emerging

Technology Trends in Electronics, Communication and Networking

(ET2ECN), 2nd International Conference on. IEEE, pp. 1-6, Dec.2014. [21] H. Kaiming, S. Jian, and T. Xiaoou, “Guided image filtering,” IEEE

transactions on pattern analysis and machine intelligence, vol. 35, no. 6,

pp. 1397-1409, Jun. 2013.

[22] Y. Song, W. Wu, Z. Liu, X. Yang, K. Liu, “An adaptive pansharpening

method by using weighted least squares filter,” IEEE Geoscience and Remote Sensing Letters, vol. 13, no. 1, pp. 18-22, Jan. 2016.

[23] Gunturk, K. Bahadir, “Fast bilateral filter with arbitrary range and

domain kernels,” IEEE Transactions on Image Processing, vol. 20, no. 9, pp. 2690-2696, Sep. 2011.

[24] F. Kou, W. Chen, C. Wen, Z. Li “Gradient domain guided image

filtering,” IEEE Transactions on Image Processing, vol. 24, no. 11, pp. 4528-4539, Nov. 2015.

[25] L. Wald, T. Ranchin, and M. Mangolini, “Fusion of satellite images of

different spatial resolutions: Assessing the quality of resulting images,” Photogrammetric Engineering Remote Sensing, vol. 63, no. 6, pp. 691–

699, Jun. 1997.

Pansharpening via Deep Guided Filtering Networkijipvc.org/article/IJIPVCV5I101.pdf · 2018. 5....

Documents

Transcript of Pansharpening via Deep Guided Filtering Networkijipvc.org/article/IJIPVCV5I101.pdf · 2018. 5....